I don't get the title. Isn't AI Pessimism not only observationally equivalent, but also actually equivalent, to AI Pessimism? I suspect that you meant for the second "AI Pessimism" in the title to be "AI Optimism."
It's possible that it's a typo, although given the content of the post, I took the title to mean:
“AI pessimism is correct, if you take into account the reference class of AI pessimism.”
In other words, if you play the game of reference class tennis in which you view your own reference class as correct and all the other references classes as incorrect, then yes, my observations are correct.
With a few exceptions at the extremes the "AI pessimists" and "AI optimists" are the SAME PEOPLE. They just differ in focus and what they were talking about when quoted (by you).
85% of them think AI is both risky and likely to produce a New Age of nearly unlimited abundance.
That 85% thinks there's some chance AI will kill us all (existential risk), and that *if it doesn't*, then things will be wonderful.
They differ on p(doom), their estimates of how likely the first is vs. the second, but virtually all of them have numbers in the range 0.1 to 0.9.
(And FWIW, none of these people think AI job loss is a real economic problem, except maybe transitionally. They worry about death, not poverty.)
“By any reasonable definition, AGI arrived a couple of years ago.”
I think I agree with most of the rest of the piece, but that “G” in the middle of AGI stands for “General”.
The idea that 2024’s LLMs had general intelligence I just don’t agree with.
I don’t think I agree with it for the LLMs we have right now in late 2026, but I acknowledge one can at least make a pretty good case.
It’s surely more than 50% we get there in the next several years, though I don’t think it’s 100% or even 98%, at least by the generally accepted definition in the industry (which to me is roughly equal to the 90th - 95th percentile human in every field, but at minimum is equal to the median human at essentially cognitive task humans do).
I'm much closer to your position than even Scott Alexander's, let alone the real doomers, but it still feels like you're a bit too dismissive.
If astronomers all came out and said there was an asteroid that would intercept Earth in March 2028, you shouldn't put them in the same reference class as every other world ending predictors and dismiss them.
It is *possible* the world gets destroyed.
The usual world-ending predictors are using magic/supernatural world-models. Those we can safely dismiss. The doomers are somewhat unique in that they're using an actual materialistic world view, presenting logical steps, and reaching a world-destroyed conclusion.
There's *probably* a logical implication they're much too certain of. It's probably along the lines of AI models having independent wants such that they'd prefer to kill or disempower humans. To paraphrase Penn Jillette, in the end, AI will probably murder as many people as it wants to murder, and that number will be zero. A ban on AI development until we have a complete understanding of how "minds" and how they can even "wants" something is probably expected value negative, but lots of attention on the potential problem such that developers slow down enough to at least ask "how many people do you *want* to murder?" with reasonable checks that they're not being deceptive is probably wise.
I don't get the title. Isn't AI Pessimism not only observationally equivalent, but also actually equivalent, to AI Pessimism? I suspect that you meant for the second "AI Pessimism" in the title to be "AI Optimism."
I had the same reaction, DRH. But the article itself is absolutely fabulous.
It's possible that it's a typo, although given the content of the post, I took the title to mean:
“AI pessimism is correct, if you take into account the reference class of AI pessimism.”
In other words, if you play the game of reference class tennis in which you view your own reference class as correct and all the other references classes as incorrect, then yes, my observations are correct.
I think you're badly misreading this Bryan.
With a few exceptions at the extremes the "AI pessimists" and "AI optimists" are the SAME PEOPLE. They just differ in focus and what they were talking about when quoted (by you).
85% of them think AI is both risky and likely to produce a New Age of nearly unlimited abundance.
That 85% thinks there's some chance AI will kill us all (existential risk), and that *if it doesn't*, then things will be wonderful.
They differ on p(doom), their estimates of how likely the first is vs. the second, but virtually all of them have numbers in the range 0.1 to 0.9.
(And FWIW, none of these people think AI job loss is a real economic problem, except maybe transitionally. They worry about death, not poverty.)
It's disingenuous to use only “human extinction” as the base rate for pessimists, instead of say “civilizational collapse and massive death”.
Many people throughout history have also predicted the latter and been wrong.
But millions of Khwarazmians, Incans, Mycenaeans, and many others predicting that would have been correct.
“By any reasonable definition, AGI arrived a couple of years ago.”
I think I agree with most of the rest of the piece, but that “G” in the middle of AGI stands for “General”.
The idea that 2024’s LLMs had general intelligence I just don’t agree with.
I don’t think I agree with it for the LLMs we have right now in late 2026, but I acknowledge one can at least make a pretty good case.
It’s surely more than 50% we get there in the next several years, though I don’t think it’s 100% or even 98%, at least by the generally accepted definition in the industry (which to me is roughly equal to the 90th - 95th percentile human in every field, but at minimum is equal to the median human at essentially cognitive task humans do).
I'm much closer to your position than even Scott Alexander's, let alone the real doomers, but it still feels like you're a bit too dismissive.
If astronomers all came out and said there was an asteroid that would intercept Earth in March 2028, you shouldn't put them in the same reference class as every other world ending predictors and dismiss them.
It is *possible* the world gets destroyed.
The usual world-ending predictors are using magic/supernatural world-models. Those we can safely dismiss. The doomers are somewhat unique in that they're using an actual materialistic world view, presenting logical steps, and reaching a world-destroyed conclusion.
There's *probably* a logical implication they're much too certain of. It's probably along the lines of AI models having independent wants such that they'd prefer to kill or disempower humans. To paraphrase Penn Jillette, in the end, AI will probably murder as many people as it wants to murder, and that number will be zero. A ban on AI development until we have a complete understanding of how "minds" and how they can even "wants" something is probably expected value negative, but lots of attention on the potential problem such that developers slow down enough to at least ask "how many people do you *want* to murder?" with reasonable checks that they're not being deceptive is probably wise.
Fascinating and absolutely convincing for me.
Looks like Bryan corrected the typo!